Repository navigation
fix(qwen4_exp): load unscaled Quark MTP weights as BF16 - #2606
Merged
Merged
Conversation
Quark lists excluded modules one by one and misses mtp.fc_embedding and mtp.layers.0.mlp.experts on amd/Qwen3.8-Flash-Next-Quark-MXFP4. The MTP head is stored entirely in BF16, with no scale tensors, but ATOM still applies the MXFP4 spec and the draft load fails. When the checkpoint index shows no MTP scale tensors, treat every mtp.* layer as excluded. A head that does carry scales, or a checkpoint with no readable index, keeps the configured policy.
Contributor
🏷️ CI GuideRuns automatically on every eligible PR before approval:
Heavy model tests:
|
This was referenced Oct 20, 2026
Merged
Merged
valarLip
approved these changes
Oct 26, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
amd/Qwen3.8-Flash-Next-Quark-MXFP4missesmtp.fc_embeddingandmtp.layers.0.mlp.experts. The whole MTP head is BF16 and has no scale tensors, but ATOM still applies the MXFP4 spec and draft loading fails.mtp.*scale tensors, resolve everymtp.*layer as excluded. A head that carries scales, or a checkpoint with no readable index, keeps the configured policy.Test plan
amd/Qwen3.8-Flash-Next-Quark-MXFP4with--method mtp --num-speculative-tokens 3and the draft weights come up.